Short-Horizon vs Long-Horizon Risk Models

A risk model estimates a covariance matrix from historical returns, but "historical" hides a critical design choice: over what window, and weighted how? A model built for a day-trading book needs to react to this week's volatility spike almost immediately, so it uses short lookback windows (weeks) and heavy exponential decay weighting that discounts old data fast. A model built for a pension fund rebalancing quarterly needs a stable, less noisy estimate of the typical covariance structure, so it uses long lookback windows (years) with much gentler decay, deliberately smoothing over short-term regime shifts.

Using a short-horizon model for a long-horizon portfolio overreacts to noise — it will flag a position as suddenly "risky" because of a one-week correlation blip that has no bearing on a multi-month holding period, causing unnecessary turnover from constant re-hedging. Using a long-horizon model for a short-horizon book does the opposite: it understates risk during a fast-moving regime change because the estimate is still anchored to a calmer multi-year average, leaving positions oversized exactly when volatility is spiking.

Worked illustration: suppose two stocks had correlation 0.3 over the past three years but correlation has jumped to 0.8 in just the last two weeks (both got caught in the same news-driven selloff). A long-horizon model still reports 0.3 and would size a pairs position as if the stocks were fairly independent; a short-horizon model would catch the 0.8 and immediately shrink recommended position size for that pair — the "correct" answer depends entirely on whether the position will actually be held for two weeks or two years.